Rotating machinery operation and maintenance data missing reconstruction method, system, medium and equipment

Abstract

A rotating machinery operation and maintenance data missing reconstruction method, system, medium and equipment, in the method, the rotating machinery multi-source operation and maintenance data with missing is collected, the rotating machinery multi-source operation and maintenance data is denoised through wavelet threshold denoising method;The denoised rotating machinery multi-source operation and maintenance data is divided by sliding window, a compressed sensing convolutional neural network model is constructed according to the size of data window and the number of data channels, a sparse representation domain matrix of rotating machinery multi-source operation and maintenance data is established, the sparse representation domain matrix is taken as the input of the compressed sensing convolutional neural network model, the reconstructed multi-source compressed data is output, the model is trained through the constructed loss function, the convolution kernel parameters in the domain coefficient convolution layer are optimized, when the model converges, the reconstructed multi-source complete data in the model training process is extracted, and the missing data reconstruction of the next data window is carried out until all the multi-source complete data is reconstructed.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery operation and maintenance technology, and in particular to a method, system, medium and equipment for reconstructing missing operation and maintenance data of rotating machinery. Background Technology

[0002] In the field of intelligent operation and maintenance of rotating machinery, a large number of different types of sensors are deployed around the core transmission components of rotating machinery for real-time monitoring. For example, sensors for vibration acceleration, sound pressure, rotational speed, and temperature are deployed around components such as gearboxes and bearings. In actual engineering practice, sudden malfunctions of sensors or data acquisition equipment can lead to the loss of operation and maintenance data. High-precision reconstruction of the missing data is necessary to ensure the accuracy of operation and maintenance analysis. Compressed sensing technology has been widely applied and researched in areas such as data acquisition and reconstruction for rotating machinery monitoring and wireless data transmission. It requires convex optimization algorithms to obtain a sparse representation of the data, but good recovery results can only be guaranteed under specific conditions.

[0003] The information disclosed in the background section is only for enhancing the understanding of the background of this invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] This invention provides a method, system, medium, and equipment for reconstructing missing operation and maintenance data of rotating machinery, which overcomes the problem of missing data in operation and maintenance. This invention can solve the synchronous reconstruction of missing operation and maintenance data from multiple sources of rotating machinery, and considers the sparse correlation between data to achieve highly reliable data reconstruction.

[0005] A method for reconstructing missing operation and maintenance data of rotating machinery includes:

[0006] Step S1: Collect multi-source operation and maintenance data of rotating machinery with missing parts, and denoise the multi-source operation and maintenance data of rotating machinery using wavelet threshold denoising method;

[0007] Step S2: Perform sliding window segmentation on the noise-reduced multi-source operation and maintenance data of rotating machinery, and construct a compressed sensing convolutional neural network model according to the size of the data window and the number of data channels. The compressed sensing convolutional neural network model includes a sparse representation domain input layer, a domain coefficient convolutional layer, an operation layer, and a compressed sampling layer.

[0008] Step S3: Establish a sparse representation domain matrix for multi-source operation and maintenance data of rotating machinery. Use the sparse representation domain matrix as input to the compressed sensing convolutional neural network model and output the reconstructed multi-source compressed data. Train the model through the constructed loss function, optimize the convolution kernel parameters in the domain coefficient convolutional layer, and when the model converges, extract the reconstructed multi-source complete data during the model training process and reconstruct the missing data of the next data window until all multi-source complete data are reconstructed.

[0009] In the aforementioned method for reconstructing missing data in rotating machinery operation and maintenance, the wavelet threshold denoising method includes:

[0010] Wavelet decomposition was performed on the multi-source operation and maintenance data of rotating machinery with missing data, and the dB10 wavelet basis was selected to decompose it into three levels of detail coefficients. , , and an approximation coefficient ,

[0011] Estimate the noise level in missing multi-source operation and maintenance data of rotating machinery by... The median is used to estimate the standard deviation of noise. Then calculate the threshold. :

[0012]

[0013]

[0014] Where N is the length of the data window in the model. This indicates taking the median.

[0015] Soft thresholding denoising was performed on the multi-source operation and maintenance data of rotating machinery with missing information, and each detailed coefficient was processed during the process. , , If its absolute value is greater than the threshold If the absolute value is less than the threshold, the coefficient is retained and a threshold correction is performed; if the absolute value is less than the threshold, it is set to zero for each wavelet coefficient. The soft thresholding denoising operation is as follows:

[0016]

[0017] in The sign function of the coefficients is expressed as:

[0018]

[0019] Detail coefficient after noise reduction , , and approximation coefficients The signal used to reconstruct the denoised signal. The reconstruction is performed by combining the inverse discrete wavelet transform (IDWT) with the denoised coefficients, as shown in the following formula:

[0020]

[0021] in It is the inverse discrete wavelet transform.

[0022] In the aforementioned method for reconstructing missing rotating machinery operation and maintenance data, before denoising the missing multi-source operation and maintenance data of rotating machinery, the missing parts are filled with 0, and after denoising, the missing parts are restored to null values. The missing multi-source operation and maintenance data of rotating machinery is divided into non-overlapping data windows, and the missing parts are filled with NaN. Data that is insufficient for the length of the data window is filled with 0.

[0023] In the aforementioned method for reconstructing missing operation and maintenance data of rotating machinery, the sparse representation domain input layer of the compressed sensing convolutional neural network model has two inputs, namely the sparse representation domain matrix. Real field matrix and imaginary field matrix Domain coefficient convolutional layers contain domain coefficients coefficients of the real field and imaginary field coefficients The intermediate hidden layer data obtained through calculation uses a linear activation function. The complete signal is obtained through the computation layer. real part and imaginary part The reconstructed compressed signal is obtained through a compressed sampling layer. real part and imaginary part N represents the number of sampling points, K represents the number of channels, and the complete signal. The specific calculation process is as follows:

[0024]

[0025]

[0026]

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] in Represents the imaginary unit, that is , This is the sampling matrix in the compressed sampling layer. For Hadama accumulation.

[0033] In the aforementioned method for reconstructing missing rotating machinery operation and maintenance data, the sparse representation domain matrix is ​​selected based on the missing multi-source operation and maintenance data of the rotating machinery. If the missing multi-source operation and maintenance data of the rotating machinery is stationary, the Discrete Fourier Complex Sparse Representation Domain Matrix (DFT) is selected; when the operation and maintenance data of the rotating machinery is non-stationary, the Harman Wavelet Complex Sparse Representation Domain Matrix is ​​selected.

[0034] The size of the discrete Fourier complex sparse representation domain matrix is Its generation process is as follows:

[0035]

[0036]

[0037] in and These are the row and column indices of the sparse representation field matrix, from 0 to... , It is a complex root of the basic unit. The imaginary unit, i.e. ,

[0038] The size of the complex sparse representation domain matrix of the Harman wavelet is Its generation process is as follows:

[0039]

[0040] in The number of wavelet subsets, It is the wavelet width.

[0041] Each wavelet subset in the Harman wavelet complex sparse representation domain matrix contains A length of All-zero column vector Wavelet subset index Index of all-zero column vectors in each wavelet First of all When assigning values, the following situations apply:

[0042]

[0043] This expression represents... The Middle arrive The data in the row is assigned the value 1.

[0044] right Performing the inverse Fourier transform yields the time-domain sequence. And normalize it to obtain This ensures that its maximum amplitude is 1, and it is stored sequentially in the Harman wavelet complex sparse representation domain matrix. The calculation process for each column is as follows:

[0045]

[0046]

[0047] in Indicates the inverse Fourier transform. This indicates finding the maximum value.

[0048] In the aforementioned method for reconstructing missing operation and maintenance data of rotating machinery, the domain coefficients are iteratively updated using random initialization. Elastic Net regularization is employed during the update process. The network's loss function L includes mean squared error and a regularization term, and the loss function is set as follows:

[0049]

[0050] in This indicates the number of channels in multi-source data. This represents the value in the original missing data of the j-th channel. This represents the value in the missing data of the reconstructed j-th channel. and They are control Regularization and Regularized weights, and These represent calculations for the 1-norm and 2-norm, respectively.

[0051] In the method for reconstructing missing operation and maintenance data of rotating machinery, the domain coefficients are randomly initialized and optimized using the Adam algorithm. The optimization process and steps are as follows:

[0052]

[0053]

[0054] in and These are the first and second moments of the gradient at the current iteration step t. and These are the attenuation factors for the first and second moments, respectively. This is the current parameter value. and These are the gradient at the current time and the square of the gradient, respectively.

[0055] In the early stages of the Adam algorithm, due to and The initialization of the gradient to zero leads to a bias in its estimated value. Therefore, it is necessary to perform bias correction to make it closer to the true gradient information. The bias correction process is as follows:

[0056] ,

[0057] in It is the first moment after bias correction. It is the second moment after bias correction; Adam uses the corrected second moment. and To update the parameters:

[0058]

[0059] in, It's the learning rate. It is a very small constant.

[0060] The Adam algorithm steps are as follows:

[0061] Step 1, Initialize parameters: Initialize the first moment Second moment Learning rate and attenuation factor and .

[0062] Step 2, Iterative Update: Calculate the current gradient Update the first-order moment and second moment ,right and Perform bias correction and use the corrected value. and Update parameters .

[0063] Step 3, Stopping condition: Repeat step 2 until the loss function converges or the predetermined maximum number of iterations is reached.

[0064] A system for implementing the method includes:

[0065] The acquisition module is used to acquire multi-source operation and maintenance data of rotating machinery with missing parts, and to reduce the noise of the multi-source operation and maintenance data of rotating machinery using the wavelet threshold denoising method.

[0066] The building unit is used to perform sliding window segmentation on the noise-reduced multi-source operation and maintenance data of rotating machinery. Based on the size of the data window and the number of data channels, a compressed sensing convolutional neural network model is constructed. The compressed sensing convolutional neural network model includes a sparse representation domain input layer, a domain coefficient convolutional layer, an operation layer, and a compressed sampling layer.

[0067] The reconstruction unit is used to establish a sparse representation domain matrix of multi-source operation and maintenance data of rotating machinery. The sparse representation domain matrix is ​​used as the input of the compressed sensing convolutional neural network model, and the output is the reconstructed multi-source compressed data. The model is trained by the constructed loss function, and the convolution kernel parameters in the domain coefficient convolutional layer are optimized. When the model converges, the reconstructed multi-source complete data during the model training process is extracted, and the missing data of the next data window is reconstructed until all multi-source complete data is reconstructed.

[0068] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0069] An electronic device, the electronic device comprising:

[0070] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0071] The processor implements the method when executing the program.

[0072] Compared with existing technologies, this invention has the following advantages: This invention performs wavelet threshold denoising on missing signals, improving signal sparsity and reducing the impact of noise on data reconstruction accuracy; the constructed compressed sensing convolutional neural network incorporates the mathematical principles of compressed sensing, making the convolutional neural network interpretable; at the same time, in the process of data missing reconstruction, the non-convex optimization algorithm of the compressed sensing convolutional neural network is used to optimize the sparse representation of the data, and the solution process combines the sparse correlation between missing data in multiple channels, realizing high-confidence reconstruction of multi-source missing data. Attached Figure Description

[0073] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0074] In the attached diagram:

[0075] Figure 1 This is a network architecture diagram of the rotating machinery operation and maintenance data missing reconstruction method of the present invention;

[0076] Figure 2 This is a schematic diagram of the overall process of the rotating machinery operation and maintenance data missing reconstruction method of the present invention;

[0077] Figure 3 This is a schematic diagram of the Harman wavelet complex sparse representation domain matrix heatmap of the rotating machinery operation and maintenance data missing reconstruction method of the present invention, wherein the selected data window is 16;

[0078] Figure 4 This is a schematic diagram comparing the reconstructed time-domain data in an embodiment of the rotating machinery operation and maintenance data missing reconstruction method of the present invention. The data is gearbox vibration data with a missing rate of 20%. For ease of display, the first 2048 sampling points are plotted.

[0079] Figure 5 This is a schematic diagram comparing the reconstructed time-domain data in an embodiment of the rotating machinery operation and maintenance data missing reconstruction method of the present invention. The data is motor torque data with a missing rate of 20%. For ease of display, the first 2048 sampling points are plotted.

[0080] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0081] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0082] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0083] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0084] like Figures 1 to 5 As shown, the method for reconstructing missing operation and maintenance data of rotating machinery includes the following steps:

[0085] Step S1: Collect multi-source operation and maintenance data of rotating machinery with missing parts, and denoise the multi-source operation and maintenance data of rotating machinery using wavelet threshold denoising method;

[0086] Step S2: Perform sliding window segmentation on the noise-reduced multi-source operation and maintenance data of rotating machinery, and construct a compressed sensing convolutional neural network model according to the size of the data window and the number of data channels. The compressed sensing convolutional neural network model includes a sparse representation domain input layer, a domain coefficient convolutional layer, an operation layer, and a compressed sampling layer.

[0087] Step S3: Establish a sparse representation domain matrix for multi-source operation and maintenance data of rotating machinery. Use the sparse representation domain matrix as input to the compressed sensing convolutional neural network model and output the reconstructed multi-source compressed data. Train the model through the constructed loss function, optimize the convolution kernel parameters in the domain coefficient convolutional layer, and when the model converges, extract the reconstructed multi-source complete data during the model training process and reconstruct the missing data of the next data window until all multi-source complete data are reconstructed.

[0088] In a preferred embodiment of the method for reconstructing missing maintenance data of rotating machinery, the wavelet threshold denoising method includes:

[0089] Wavelet decomposition was performed on the multi-source operation and maintenance data of rotating machinery with missing data, and the dB10 wavelet basis was selected to decompose it into three levels of detail coefficients. , , and an approximation coefficient ,

[0090] Estimate the noise level in missing multi-source operation and maintenance data of rotating machinery by... The median is used to estimate the standard deviation of noise. Then calculate the threshold. :

[0091]

[0092]

[0093] Where N is the length of the data window in the model. This indicates taking the median.

[0094] Soft thresholding denoising was performed on the multi-source operation and maintenance data of rotating machinery with missing information, and each detailed coefficient was processed during the process. , , If its absolute value is greater than the threshold If the absolute value is less than the threshold, the coefficient is retained and a threshold correction is performed; if the absolute value is less than the threshold, it is set to zero for each wavelet coefficient. The soft thresholding denoising operation is as follows:

[0095]

[0096] in The sign function of the coefficients is expressed as:

[0097]

[0098] Detail coefficient after noise reduction , , and approximation coefficients The signal used to reconstruct the denoised signal. The reconstruction is performed by combining the inverse discrete wavelet transform (IDWT) with the denoised coefficients, as shown in the following formula:

[0099]

[0100] in It is the inverse discrete wavelet transform.

[0101] In a preferred embodiment of the method for reconstructing missing rotating machinery operation and maintenance data, before denoising the missing multi-source operation and maintenance data of rotating machinery, the missing parts are filled with 0, and after denoising, the missing parts are restored to null values. The missing multi-source operation and maintenance data of rotating machinery is divided into non-overlapping data windows, and the missing parts are occupied by NaN. Data that is insufficient for the length of the data window is filled with 0.

[0102] In a preferred embodiment of the method for reconstructing missing operation and maintenance data of rotating machinery, the sparse representation domain input layer in the compressed sensing convolutional neural network model has two inputs, namely the sparse representation domain matrix. Real field matrix and imaginary field matrix Domain coefficient convolutional layers contain domain coefficients coefficients of the real field and imaginary field coefficients The intermediate hidden layer data obtained through calculation uses a linear activation function. The complete signal is obtained through the computation layer. real part and imaginary part The reconstructed compressed signal is obtained through a compressed sampling layer. real part and imaginary part N represents the number of sampling points, K represents the number of channels, and the complete signal. The specific calculation process is as follows:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111] in Represents the imaginary unit, that is , This is the sampling matrix in the compressed sampling layer. For Hadama accumulation.

[0112] In a preferred embodiment of the method for reconstructing missing rotating machinery operation and maintenance data, the sparse representation domain matrix is ​​selected based on the missing multi-source operation and maintenance data of the rotating machinery. If the missing multi-source operation and maintenance data of the rotating machinery is stationary, the Discrete Fourier Complex Sparse Representation Domain Matrix (DFT) is selected; if the operation and maintenance data of the rotating machinery is non-stationary, the Harman Wavelet Complex Sparse Representation Domain Matrix is ​​selected.

[0113] The size of the discrete Fourier complex sparse representation domain matrix is Its generation process is as follows:

[0114]

[0115]

[0116] in and These are the row and column indices of the sparse representation field matrix, from 0 to... , It is a complex root of the basic unit. The imaginary unit, i.e. ,

[0117] The size of the complex sparse representation domain matrix of the Harman wavelet is Its generation process is as follows:

[0118]

[0119] in The number of wavelet subsets, It is the wavelet width.

[0120] Each wavelet subset in the Harman wavelet complex sparse representation domain matrix contains A length of All-zero column vector Wavelet subset index Index of all-zero column vectors in each wavelet First of all When assigning values, the following situations apply:

[0121]

[0122] This expression represents... The Middle arrive The data in the row is assigned the value 1.

[0123] right Performing the inverse Fourier transform yields the time-domain sequence. And normalize it to obtain This ensures that its maximum amplitude is 1, and it is stored sequentially in the Harman wavelet complex sparse representation domain matrix. The calculation process for each column is as follows:

[0124]

[0125]

[0126] in Indicates the inverse Fourier transform. This indicates finding the maximum value.

[0127] In a preferred embodiment of the method for reconstructing missing operation and maintenance data of rotating machinery, the domain coefficients are iteratively updated using random initialization. Elastic Net regularization is employed during the update process. The network's loss function L includes mean squared error and a regularization term, and the loss function is set as follows:

[0128]

[0129] in This indicates the number of channels in multi-source data. This represents the value in the original missing data of the j-th channel. This represents the value in the missing data of the reconstructed j-th channel. and They are control Regularization and Regularized weights, and These represent calculations for the 1-norm and 2-norm, respectively.

[0130] In a preferred embodiment of the method for reconstructing missing operation and maintenance data of rotating machinery, the domain coefficients are randomly initialized and optimized using the Adam algorithm. The optimization process and steps are as follows:

[0131]

[0132]

[0133] in and These are the first and second moments of the gradient at the current iteration step t. and These are the attenuation factors for the first and second moments, respectively. This is the current parameter value. and These are the gradient at the current time and the square of the gradient, respectively.

[0134] In the early stages of the Adam algorithm, due to and The initialization of the gradient to zero leads to a bias in its estimated value. Therefore, it is necessary to perform bias correction to make it closer to the true gradient information. The bias correction process is as follows:

[0135] ,

[0136] in It is the first moment after bias correction. It is the second moment after bias correction; Adam uses the corrected second moment. and To update the parameters:

[0137]

[0138] in, It's the learning rate. It is a very small constant.

[0139] The Adam algorithm steps are as follows:

[0140] Step 1, Initialize parameters: Initialize the first moment Second moment Learning rate and attenuation factor and .

[0141] Step 2, Iterative Update: Calculate the current gradient Update the first-order moment and second moment ,right and Perform bias correction and use the corrected value. and Update parameters .

[0142] Step 3, Stopping condition: Repeat step 2 until the loss function converges or the predetermined maximum number of iterations is reached.

[0143] A system for implementing the method includes:

[0144] The acquisition module is used to acquire multi-source operation and maintenance data of rotating machinery with missing parts, and to reduce the noise of the multi-source operation and maintenance data of rotating machinery using the wavelet threshold denoising method.

[0145] The building unit is used to perform sliding window segmentation on the noise-reduced multi-source operation and maintenance data of rotating machinery. Based on the size of the data window and the number of data channels, a compressed sensing convolutional neural network model is constructed. The compressed sensing convolutional neural network model includes a sparse representation domain input layer, a domain coefficient convolutional layer, a complex number operation layer, and a compressed sampling layer.

[0146] The reconstruction unit is used to establish a sparse representation domain matrix of multi-source operation and maintenance data of rotating machinery. The sparse representation domain matrix is ​​used as the input of the compressed sensing convolutional neural network model, and the output is the reconstructed multi-source compressed data. The model is trained by the constructed loss function, and the convolution kernel parameters in the domain coefficient convolutional layer are optimized. When the model converges, the reconstructed multi-source complete data during the model training process is extracted, and the missing data of the next data window is reconstructed until all multi-source complete data is reconstructed.

[0147] In one embodiment, such as Figure 2 As shown, to address the problem of missing multi-source operation and maintenance data during the intelligent operation and maintenance of rotating machinery, this invention proposes a method for reconstructing missing operation and maintenance data for rotating machinery. First, wavelet threshold denoising is applied to the missing multi-source data to reduce the impact of noise on the sparse representation of the missing data. Then, the missing data is divided into data windows and input into a compressed sensing convolutional neural network model in batches for training. By constructing a sparse representation domain matrix and a sampling matrix, the domain coefficients are continuously optimized to approximate the features of the original complete data, thereby achieving the reconstruction of the missing multi-source data.

[0148] The specific method for reconstructing missing multi-source operation and maintenance data of rotating machinery based on compressed sensing convolutional neural networks includes the following steps:

[0149] Step S1: Load the multi-source operation and maintenance data of the rotating machinery with missing parts. In this specific embodiment, the gearbox tooth breakage data of Southeast University is used. The loaded data is denoised by wavelet threshold denoising method.

[0150] Specifically, before denoising, the missing parts are filled with 0, and after denoising, the missing parts are restored to null values. The wavelet thresholding denoising method is as follows:

[0151] Wavelet decomposition was performed on the missing data, and the dB10 wavelet basis was selected to decompose it into three levels of detail coefficients. , , ) and an approximation coefficient .

[0152] Estimate the noise level in the missing data by The median is used to estimate the standard deviation of noise. Then calculate the threshold. :

[0153]

[0154]

[0155] Where N is the length of the data window in the model. This indicates taking the median.

[0156] Soft thresholding is used to denoise the missing data, and each detail coefficient is processed during the process. , , If its absolute value is greater than the threshold If the absolute value is less than the threshold, the coefficient is retained and a threshold correction is applied; if the absolute value is less than the threshold, it is set to zero. For each wavelet coefficient... The soft thresholding denoising operation is as follows:

[0157]

[0158] in The sign function of the coefficients is expressed as:

[0159]

[0160] Detail coefficient after noise reduction ( , , ) and approximation coefficients The signal used to reconstruct the denoised signal. Reconstruction is performed using the inverse discrete wavelet transform (IDWT) combined with the denoised coefficients, as shown in the formula:

[0161]

[0162] in It is the inverse discrete wavelet transform.

[0163] In this specific embodiment, complete data is used, and noise is removed by wavelet thresholding. The denoised data is then used to generate multi-source missing data through a sampling matrix, which facilitates comparison with the generated complete data. In engineering practice, multi-source missing data can be used. The missing data is divided into non-overlapping data windows of length 2048, with NaN used as placeholders for the missing parts, and zeros are used to pad any data portions less than 2048.

[0164] Step S2: Perform a sliding window segmentation of length 2048 on the denoised missing data. Assuming it is 6-channel multi-source data, construct a compressed sensing convolutional neural network model (e.g., Figure 1 As shown, the sparse representation domain includes a sparse representation input layer (2*2048*2048, this layer is a two-layer input structure), a domain coefficient convolutional layer (convolution kernel size is 6*2048*1, a total of four convolution kernels), an operation layer (2048*6), and a compressed sampling layer (2048*6). The real field matrix and the imaginary field matrix of the sparse representation domain each have two convolution kernels, and these two convolution kernels are the real field coefficients. and imaginary field coefficients ;

[0165] Specifically, in the compressed sensing convolutional neural network model, the sparse representation domain input layer has two inputs: the sparse representation domain matrix and the input layer. Real field matrix and imaginary field matrix Domain coefficient convolutional layers contain domain coefficients coefficients of the real field and imaginary field coefficients The calculated intermediate hidden layer data, in the convolutional neural network computation, has a kernel size of 2048*1 for each channel, with 6 channels, and the activation function is a linear function. The complete signal is obtained through the computation layer. real part and imaginary part The reconstructed compressed signal is obtained through a compressed sampling layer. real part and imaginary part The number of sampling points is 2048, the number of channels is 6, and the complete signal is... The specific calculation process is as follows:

[0166]

[0167]

[0168]

[0169]

[0170]

[0171]

[0172]

[0173]

[0174] in Represents the imaginary unit, that is , This is the sampling matrix in the compressed sampling layer. For Hadama accumulation.

[0175] Step S3: Based on the stationarity of the data, a sparse representation domain matrix for the multi-source missing data is established. This sparse representation domain matrix is ​​used as the input to the compressed sensing convolutional neural network model, outputting reconstructed multi-source compressed data. The model is trained using the constructed loss function, optimizing the convolution kernel parameters in the domain coefficient convolutional layer. When the model converges, the reconstructed multi-source complete data obtained during model training is extracted. Then, reconstruct the missing data for the next data window;

[0176] Specifically, the sparse domain matrix can be selected based on the characteristics of the missing data. If the rotating machinery operation and maintenance data is stationary, the Discrete Fourier Complex Sparse Domain Matrix (DFT) can be used; if the rotating machinery operation and maintenance data is non-stationary, the Harman Wavelet Complex Sparse Domain Matrix can be used. In practical engineering, the healthy operation data of rotating machinery is stationary, while the fault data is non-stationary. This specific implementation uses these two types of data as examples to illustrate the generation process of the Discrete Fourier Complex Sparse Domain Matrix and the Harman Wavelet Complex Sparse Domain Matrix:

[0177] The size of the discrete Fourier complex sparse representation domain matrix is Its generation process is as follows:

[0178]

[0179]

[0180] in and These are the row and column indices of the sparse representation domain matrix (from 0 to 2048). It is a complex root of the basic unit. The imaginary unit (i.e.) ).

[0181] The size of the complex sparse representation domain matrix of the Harman wavelet is The wavelet width is set to 2, and the number of wavelet subsets is 1024. The generation process is as follows:

[0182] Each wavelet subset in the Harman wavelet complex sparse representation domain matrix contains two all-zero column vectors of length 2048. Wavelet subset index Index of all-zero column vectors in each wavelet First of all When assigning values, the following situations apply:

[0183]

[0184] This expression represents... The Middle arrive The data in the row is assigned the value 1.

[0185] right Performing the inverse Fourier transform (IFFT) yields the time-domain sequence. And normalize it to obtain This ensures that its maximum amplitude is 1, and it is stored sequentially in the Harman wavelet complex sparse representation domain matrix. The calculation process for each column is as follows:

[0186]

[0187]

[0188] in Indicates the inverse Fourier transform. This indicates finding the maximum value.

[0189] Specifically, the domain coefficients are iteratively updated using random initialization, and ElasticNet regularization is applied during the update process. The network's loss function includes mean squared error and a regularization term, and the loss function L is set as follows:

[0190]

[0191] in This indicates the number of channels for multi-source data. In this specific implementation, 6 channels of data are selected. This represents the value in the original missing data of the j-th channel. This represents the value in the missing data of the reconstructed j-th channel. and They are control Regularization and The weights for regularization are all 10 in this specific implementation.

[0192] Specifically, the field coefficients are randomly initialized and optimized using the Adam algorithm. The optimization process and steps are as follows:

[0193]

[0194]

[0195] in and These are the first and second moments of the gradient at the current iteration step t. and These are the attenuation factors for the first and second moments, respectively. This is the current parameter value. and These are the gradient at the current time and the square of the gradient, respectively.

[0196] In the early stages of the Adam algorithm, due to and The initialization of the gradient to zero leads to a bias in its estimated value. Therefore, a bias correction is needed to make it closer to the true gradient information. The bias correction process is as follows:

[0197]

[0198] in It is the first moment after bias correction. It is the second moment after bias correction; Adam uses the corrected second moment. and To update the parameters:

[0199]

[0200] in, This is the learning rate, set to 0.0001. It is a very small constant.

[0201] The Adam algorithm steps are as follows:

[0202] (1) Initialize parameters: Initialize the first moment Second moment Learning rate and attenuation factor and .

[0203] (2) Iterative update: Calculate the current gradient Update the first-order moment and second moment ,right and Perform bias correction and use the corrected value. and Update parameters .

[0204] (3) Stopping condition: Repeat step (2) until the loss function converges or the predetermined maximum number of iterations is reached.

[0205] Specifically, the sampling matrix in the compressed sampling layer Following a Bernoulli distribution, the missing data points are initially marked as 0 in the matrix, and then the markers are increased according to the sampling rate. The requirement is that the number of unsampled data points is greater than the total number of missing data points. The sampling matrix construction process is as follows:

[0206]

[0207] in The sampling rate can be selected based on the missing data. Since most rotating machinery maintenance data is not sparse, the sampling rate needs to be as close as possible to the actual data to improve the accuracy of missing data reconstruction and recovery. Sampling matrix The size is , and These are the row and column indices of the sampling matrix.

[0208] See the complete extracted data Figure 4 and Figure 5 The data processed was discretely missing (missing rate 0.2%), with the shaded area representing the missing data. Multiple metrics were used to evaluate the reconstruction effect of the missing data (see Table 1). The calculation process for each metric is as follows:

[0209] Mean square error :

[0210]

[0211] Pearson correlation coefficient :

[0212]

[0213] Mean Absolute Error :

[0214]

[0215] Coefficient of determination :

[0216]

[0217] in This represents data from one channel within the original, complete, multi-source operations and maintenance data. This represents the complete data after reconstruction of missing data. express Norm, This indicates that calculations are performed column by column. and These represent the original data and reconstructed data of the i-th point in a certain channel, respectively. and These represent the average of the original data and the reconstructed data for a certain channel, respectively.

[0218] The above is only a description of a method for reconstructing missing multi-source operation and maintenance data of rotating machinery based on compressed sensing convolutional neural networks, but the present invention is not limited to this application scope.

[0219] Table 1 Evaluation of Missing Data Reconstruction Results

[0220]

[0221] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for reconstructing missing operation and maintenance data of rotating machinery, characterized in that, Includes the following steps: Step S1: Collect multi-source operation and maintenance data of rotating machinery with missing parts, and denoise the multi-source operation and maintenance data of rotating machinery using wavelet threshold denoising method; Step S2: Perform sliding window segmentation on the noise-reduced multi-source operation and maintenance data of rotating machinery, and construct a compressed sensing convolutional neural network model according to the size of the data window and the number of data channels. The compressed sensing convolutional neural network model includes a sparse representation domain input layer, a domain coefficient convolutional layer, an operation layer, and a compressed sampling layer. Step S3: Establish a sparse representation domain matrix for multi-source operation and maintenance data of rotating machinery. Use the sparse representation domain matrix as input to the compressed sensing convolutional neural network model and output the reconstructed multi-source compressed data. Train the model through the constructed loss function, optimize the convolution kernel parameters in the domain coefficient convolutional layer, and when the model converges, extract the reconstructed multi-source complete data during the model training process and reconstruct the missing data of the next data window until all multi-source complete data are reconstructed.

2. The method for reconstructing missing operation and maintenance data of rotating machinery according to claim 1, characterized in that, Preferred wavelet thresholding denoising methods include: Wavelet decomposition was performed on the multi-source operation and maintenance data of rotating machinery with missing data, and the dB10 wavelet basis was selected to decompose it into three levels of detail coefficients. , , and an approximation coefficient , Estimate the noise level in missing multi-source operation and maintenance data of rotating machinery by... The median is used to estimate the standard deviation of noise. Then calculate the threshold. : ， ， Where N is the length of the data window in the model. This indicates taking the median. Soft thresholding denoising was performed on the multi-source operation and maintenance data of rotating machinery with missing information, and each detailed coefficient was processed during the process. , , If its absolute value is greater than the threshold If the absolute value is less than the threshold, the coefficient is retained and a threshold correction is performed; if the absolute value is less than the threshold, it is set to zero for each wavelet coefficient. The soft thresholding denoising operation is as follows: ， in The sign function of the coefficients is expressed as: ， Detail coefficient after noise reduction , , and approximation coefficients The signal used to reconstruct the denoised signal. The reconstruction is performed by combining the inverse discrete wavelet transform (IDWT) with the denoised coefficients, as shown in the following formula: ， in It is the inverse discrete wavelet transform.

3. The method for reconstructing missing operation and maintenance data of rotating machinery according to claim 1, characterized in that, Before denoising the multi-source operation and maintenance data of rotating machinery with missing parts, the missing parts are filled with 0. After denoising, the missing parts are restored to null values. The multi-source operation and maintenance data of rotating machinery with missing parts are divided into non-overlapping data windows. The missing parts are filled with NaN. Data that is not long enough to fill the data window is filled with 0.

4. The method for reconstructing missing operation and maintenance data of rotating machinery according to claim 1, characterized in that, In the compressed sensing convolutional neural network model, the sparse representation domain input layer has two inputs: the sparse representation domain matrix and the input layer. Real field matrix and imaginary field matrix Domain coefficient convolutional layers contain domain coefficients coefficients of the real field and imaginary field coefficients The intermediate hidden layer data obtained through calculation uses a linear activation function. The complete signal is obtained through the computation layer. real part and imaginary part The reconstructed compressed signal is obtained through a compressed sampling layer. real part and imaginary part N represents the number of sampling points, K represents the number of channels, and the complete signal The specific calculation process is as follows: ， ， ， ， ， ， ， ， in Represents the imaginary unit, that is , This is the sampling matrix in the compressed sampling layer. For Hadama accumulation.

5. The method for reconstructing missing operation and maintenance data of rotating machinery according to claim 1, characterized in that, The sparse representation domain matrix is ​​selected based on the missing multi-source operation and maintenance data of the rotating machinery. If the missing multi-source operation and maintenance data of the rotating machinery is stationary, the Discrete Fourier Complex Sparse Representation Domain Matrix (DFT) is used; if the operation and maintenance data of the rotating machinery is non-stationary, the Harman Wavelet Complex Sparse Representation Domain Matrix is ​​used. The size of the discrete Fourier complex sparse representation domain matrix is Its generation process is as follows: ， ， in and These are the row and column indices of the sparse representation field matrix, from 0 to... , It is a complex root of the basic unit. The imaginary unit, i.e. , The size of the complex sparse representation domain matrix of the Harman wavelet is Its generation process is as follows: ， in The number of wavelet subsets, It is the wavelet width. Each wavelet subset in the Harman wavelet complex sparse representation domain matrix contains A length of All-zero column vector Wavelet subset index Index of all-zero column vectors in each wavelet First of all When assigning values, the following situations apply: ， This expression represents... The Middle arrive The data in the row is assigned the value 1. right Performing the inverse Fourier transform yields the time-domain sequence. And normalize it to obtain This ensures that its maximum amplitude is 1, and it is stored sequentially in the Harman wavelet complex sparse representation domain matrix. The calculation process for each column is as follows: ， ， in Indicates the inverse Fourier transform. This indicates finding the maximum value.

6. The method for reconstructing missing operation and maintenance data of rotating machinery according to claim 1, characterized in that, The domain coefficients are iteratively updated using random initialization, and Elastic Net regularization is applied during the update process. The network's loss function includes mean squared error and a regularization term. The settings are as follows: ， in This indicates the number of channels in multi-source data. This represents the value in the original missing data of the j-th channel. This represents the value in the missing data of the reconstructed j-th channel. and These are control Regularization and Regularized weights, and These represent calculations for the 1-norm and 2-norm, respectively.

7. The method for reconstructing missing operation and maintenance data of rotating machinery according to claim 1, characterized in that, The domain coefficients are randomly initialized and optimized using the Adam algorithm. The optimization process and steps are as follows: ， ， in and These are the first and second moments of the gradient at the current iteration step t. and These are the attenuation factors for the first and second moments, respectively. This is the current parameter value. and These are the gradient at the current time and the square of the gradient, respectively. In the early stages of the Adam algorithm, due to and The initialization of the gradient to zero leads to a bias in its estimated value. A bias correction is performed to make it closer to the true gradient information. The bias correction process is as follows: ， in It is the first moment after bias correction. It is the second moment after bias correction; Adam uses the corrected second moment. and To update the parameters: ， in, It's the learning rate. It is a very small constant. The Adam algorithm steps are as follows: Step 1, Initialize parameters: Initialize first-order moments Second moment Learning rate and attenuation factor and , Step 2, Iterative Update: Calculate the current gradient Update the first-order moment and second moment ,right and Perform bias correction and use the corrected value. and Update parameters , Step 3, Stopping condition: Repeat step 2 until the loss function converges or the predetermined maximum number of iterations is reached.

8. A system for implementing the method of any one of claims 1-7, characterized in that, It includes: The acquisition module is used to acquire multi-source operation and maintenance data of rotating machinery with missing parts, and to reduce the noise of the multi-source operation and maintenance data of rotating machinery using the wavelet threshold denoising method. The building unit is used to perform sliding window segmentation on the noise-reduced multi-source operation and maintenance data of rotating machinery. Based on the size of the data window and the number of data channels, a compressed sensing convolutional neural network model is constructed. The compressed sensing convolutional neural network model includes a sparse representation domain input layer, a domain coefficient convolutional layer, an operation layer, and a compressed sampling layer. The reconstruction unit is used to establish a sparse representation domain matrix of multi-source operation and maintenance data of rotating machinery. The sparse representation domain matrix is ​​used as the input of the compressed sensing convolutional neural network model, and the output is the reconstructed multi-source compressed data. The model is trained by the constructed loss function, and the convolution kernel parameters in the domain coefficient convolutional layer are optimized. When the model converges, the reconstructed multi-source complete data during the model training process is extracted, and the missing data of the next data window is reconstructed until all multi-source complete data is reconstructed.

9. A computer storage medium, characterized in that, The storage medium includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements the method as described in any one of claims 1-7.

Images